Autonomous Agent on Telegram

Yes, if you want an AI agent that runs 24/7, listens to your Telegram, and does work on its own — intel scans, content drafts, code commits, whatever you tell it. Once set up, you message it like a person. It replies like a person. Except it never sleeps.

1. step 1 / 6

   Install the framework

   Generate install command below → copy → Terminal → enter. The kit lands in ~/.claude/skills/autonomous-agent-telegram/ and includes the bus scripts (send, receive, restart), the fast-checker daemon, a minimal agent config template, and a launchd plist for always-on operation.

2. step 2 / 6

   Create a Telegram bot with @BotFather

   Open Telegram. Search for @BotFather. Send /newbot. Give it a name + username. BotFather returns a bot token (looks like 8309826713:AAE...). Save this — you'll paste it next step.

   Also grab your personal chat ID: message @userinfobot and send anything — it replies with your numeric ID.

3. step 3 / 6

   Name your agent + wire the token

   In Claude Code, type:

   Use autonomous-agent-telegram to scaffold a new agent called [pick-a-name] for project [pick-a-project]
   

   Example: scaffold a new agent called 'scout' for project 'my-newsletter'. The skill creates a folder, writes the CLAUDE.md instructions file, and prompts you for the bot token + chat ID. Paste them. Done.

4. step 4 / 6

   Start the fast-checker daemon

   This is the piece that makes the agent LIVE. The daemon polls Telegram every few seconds for new messages and funnels them into the agent's running session:

   bash ~/.claude/skills/autonomous-agent-telegram/scripts/start-daemon.sh [your-agent-name]
   

   You'll see a console tail. Leave it running in a terminal window (or install the launchd plist to auto-start at login — instructions in the skill's README).

5. step 5 / 6

   Talk to your agent

   Open Telegram. Find your bot (the username you gave BotFather). Send a message: hey scout, what's up? Within 1-5 seconds, your agent responds. It has context of its project via the CLAUDE.md file. Talk to it like a person.

   Try: summarize my latest git commits, pull the top 10 tweets from my X feed and tell me what matters, build me a landing page for [topic] and commit when done.

6. step 6 / 6

   Give it crons (optional, transforms the agent from reactive to proactive)

   Edit ~/your-project/agent-name/config.json → add a crons array:

   "crons": [
     { "name": "morning-digest", "interval": "0 7 * * *", "prompt": "Scan my X feed, send me the top 5 signals by 7:15." },
     { "name": "daily-build", "interval": "0 14 * * *", "prompt": "Pick the next task from TASKS.md, build it, commit, ping me." }
   ]
   

   Save. The agent auto-schedules at every fire. Now it's not just reactive — it initiates.

• Newsletter writer? Agent scans industry news daily, pulls top 5 signals, drafts a newsletter section you review+polish. Cuts 2 hours off your weekly writing.
• Indie hacker? Agent auto-commits small fixes from your TODO.md on a cron, pings you on Telegram with the diff. Sleep-shipping.
• Creator? Agent drafts 3 content ideas every morning based on yesterday's trending tweets. You pick one, it's executed by another agent (see module 12 — the build loop).
• Agency operator? One agent per client — each has its own CLAUDE.md with the client's brand voice + constraints. They handle routine Q&A + first drafts in the client's own Telegram.

• Bot doesn't respond when I message it.

  Daemon isn't running. `bash ~/.claude/skills/autonomous-agent-telegram/scripts/start-daemon.sh [agent-name]` in a terminal. Leave the window open. If you want always-on, load the launchd plist from the skill's README.

• Agent responds but context-drifts — forgets what project it's on.

  You didn't fill out the CLAUDE.md properly. Open it and write a clear 'who you are / what you work on / what's the priority' — 10 lines is enough. Agents without a clear CLAUDE.md wander.

• Token exposed in a public repo.

  Rotate immediately. Go to @BotFather → `/revoke` → old token is dead, generate new one. The `.env` files with tokens should be in `.gitignore` from the start. Never check tokens into git.

• Crons fire at weird times / not at all.

  Session-scoped crons only fire while the agent's session is open. For true always-on scheduling, use the bundled launchd plist (Mac) or crontab setup — both covered in the skill's README.

• Agent uses too many Claude credits.

  Claude Pro plan has daily caps. For long-running agents use Max plan OR install RTK (Rust Token Killer — search GitHub for `rust-token-killer`) as a CLI proxy — 60-90% token reduction on common commands. Max plan + RTK = near-unlimited agent runtime.

• ★ required
  AI agency generalist

  Bespoke AI agents and automations for mid-size companies. Higher-touch, higher-ticket cousin of the local-biz play.

• ★ required
  AI companion / character-as-a-service

  Niche AI companion w/ persistent memory + voice. Character.ai is saturated; specific niches still wide open.

• ★ required
  Claude Code for local businesses

  Install Claude Code at a plumber, lawyer, or clinic. Build them custom workflows. Hourly retainer + monthly fee.

• ★ required
  Content-at-scale for founders

  Ghostwrite + clip founders' content into a high-volume LinkedIn / X / podcast presence. Premium DFY play.

• ★ required
  Creator memberships / paid Telegram channels

  Lock your most useful daily output behind $20-50/mo. Telegram + Skool + Whop are the cheapest delivery rails.

• ★ required
  Faceless SaaS micro-app

  Find a sharp insecurity or repetitive pain. Wrap it in a single-purpose AI app. Charge weekly. Stack 5-10 of them.